Closing with Claude: use it as the opposition

The inversion

Ask a model to “assess this deal” and it will summarise your notes back to you with a tidy risk section at the bottom. That feels like analysis. It is not. Your notes were written by the person who wants the deal to close, so a summary of them is optimism with better formatting.

Ask it to argue the deal will not close and the task changes. Now the model has to find evidence, and the only place to find it is the silence in your notes: the person who has not spoken since the demo, the budget line nobody confirmed, the security review that appeared once in an email and never again. You will disagree with half of what comes back. The half you cannot argue with is the forecast risk you were carrying anyway.

Six prompts for the end of a deal

Replace anything in braces before running. The note under each one explains which line is doing the work, because that is the part to copy into your own prompts, not the wording.

Below is everything I have on an open deal: call
notes, email thread, the proposal, and the CRM fields.
{paste}

Do not assess this deal. Argue the case that it will
not close this quarter. Be specific and use only what
is in the material.

Then finish with exactly two things:
1. The single most important fact I do not have
2. The exact question I should ask, in one sentence,
   and who I should ask it to

Here are the transcripts and notes from all four
calls with this account:
{paste}

Pull out every sentence where someone expressed doubt,
hesitation, or a competing priority. Quote them exactly.
Do not paraphrase and do not clean up the wording.

Group them by the person who said it. For each group,
tell me what that person is actually worried about,
and mark it [inferred] where you are reading between
the lines rather than quoting.

Last: which objection has never been answered on a
call, only in email?

We are in a competitive evaluation against {competitor}.
Here is our product documentation and here is what I
know about theirs:
{paste}

Name the three areas where they are genuinely better
for this buyer, given what this buyer told us they
care about: {paste requirements}.

For each one, tell me:
- whether it is a real gap or a difference in approach
- what I should say if it comes up
- whether it is worth raising first

Do not give me a rebuttal for something they actually
do better. Say so plainly instead.

Build a mutual action plan to a signature date of
{date}, working backwards, using only the notes below.
{paste}

For every step include: owner, date, and what has to
be true before it can start.

Critical rule: tag anything you inferred rather than
read with [assumed]. That includes procurement time,
who signs, whether legal review is needed, and any
date I did not state.

At the end, list the [assumed] items as questions I
can send to my champion in one email.

Here are our security and compliance documents:
{paste policy, SOC 2 summary, DPA, sub-processor list}

Here are 40 questions from the buyer's security team:
{paste}

Answer only what is directly supported by the documents
above. Quote or cite the source line for each answer.

If a question is not covered, or is only partly covered,
write NEEDS INPUT and name who internally should answer
it. Do not infer, do not generalise from a similar
control, and do not soften a no into a partial yes.

Return a table: question, answer or NEEDS INPUT, source.

You are the procurement lead at {company}. You have
been handed this proposal and told to reduce spend.
{paste proposal and pricing}

Here is what I know about their process, timeline and
alternatives: {paste}

Write the three things you would push on first, in the
order you would push, and the leverage you would use
for each.

Then, as yourself again: for each one, what is the
smallest concession that resolves it, and what should
I ask for in return?

Multithreading: find out who actually decides

Most deals that slip do so because one person was carrying it internally and nobody checked whether they could. You do not need a new framework for this. You need a list of everyone who has appeared in the thread and an honest read on which of them has never spoken to you directly.

Give the model the full email history and every call note, then ask for a simple table: name, role, what they have said in their own words, last contact date, and whether you have ever spoken to them live. The people with a blank in the last column are the deal. If the CFO exists only as a cc line and a second-hand quote from your champion, you have a one-threaded deal regardless of how many logos are on the account plan. Pulling that history together is much easier when the model can read the record itself rather than waiting for you to paste it, which is what connecting Claude to your CRM is for.

The line you do not cross

Never let a model approximate a claim about security, compliance, legal terms, or pricing. Not a rounded number, not a “broadly equivalent” control, not a best guess at whether you support a data residency requirement.

The reason is not accuracy in the abstract. It is that these answers are contractual. A wrong sentence in a security questionnaire, a retention period you invented, a discount you implied you could authorise: being wrong there is a legal problem, not a productivity problem, and it lands on your company rather than on you fixing a draft. Everywhere else the cost of a bad model output is a few minutes of editing. Here it is not.

So the working rule is simple. If a claim would appear in a contract, a security review, or an invoice, the model may only repeat what is in a document you already trust, and it must cite which one. Anything else gets marked NEEDS INPUT and goes to the person whose job it is. If you are writing team guardrails, that distinction between drafting work and stateable facts is the one worth writing down first, and the AI sales checklist covers the rest of the policy surface.

Honest forecasting is mostly a data problem

Ask the model to categorise each open deal by what is missing rather than by percentage. Three buckets do more work than a probability field: no confirmed economic buyer, no confirmed process, no confirmed date. A deal missing all three is not at 60 percent, whatever the CRM says.

Then ask for the single question that would move each deal out of its bucket, and send those questions this week. That is the whole exercise. The value is not the reclassification, it is the twelve specific questions you now have to ask, most of which you had been quietly avoiding.

What this does not do

AI does not close deals. It removes the reasons you were not ready: the prep you skipped, the objection you had not written an answer to, the questionnaire that sat for nine days, the plan you never wrote down. Those are real costs and getting them back is worth a lot.

The rest is still yours. Whether a buyer trusts you, whether this is the right week to push, whether the champion is genuinely a champion or just polite, whether to walk away. No model has seen the pause before someone answered your question. Judgment, trust and timing do not transfer, and any tool that claims otherwise is selling you something. If you want the wider picture of where this fits across a full cycle, start with how sales teams actually use Claude day to day.

Then the loop starts again. The account you just closed is the template for the next twenty you go after, so take the reasons this one worked and feed them back into how you pick and open the next set.